MET Norway has recently developed Bris, a probabilistic data-driven weather forecasting model based on graph neural networks. Bris is a global model with a flexible stretched grid, allowing higher spatial resolution for specific focus areas, while maintaining a lower resolution elsewhere. Based on a stochastic encoder–decoder architecture, the model is trained using a loss function based on the Continuous Ranked Probability Score (CRPS) evaluated point-wise in real and spectral space. We will present the Bris model and verification results against MET Norway’s state-of-the-art forecasting systems, focusing on variables most relevant for public weather services, and highlight both strengths and current limitations. We will also discuss particular challenges when implementing a purely data-driven model for operational use on Yr.
Preprints are available at https://doi.org/10.48550/arXiv.2409.02891 and https://arxiv.org/html/2511.23043v1.